Basically Tiny Language Models (below 10m parameters or only one transformer block) can generate paragraphs of coherent text and reason....provided training is limited to stories that only contain words typical 3 to 4 year olds would understand.